Red Hat AI 3.5: governing AI agents in production
Red Hat AI 3.5 brings together safety evaluation, tenant isolation, agent observability, and resource metering. For Belgian and French organizations, the challenge is no longer just proving a pilot: AI must be operated as a governed, measurable, mission-critical service.
1. What Red Hat announced
Announced as generally available on September 9, the release extends EvalHub to models brought or customized by customers, with tests that cover prompt injection, jailbreaks, personal-data exposure, and toxicity. It also adds validated models with safety scores and controlled rollouts to reduce risk during model updates.
For operations, Red Hat documents per-user token metering, model and agent performance, MLflow agent tracing, and GPU utilization. Dedicated control planes and OpenShift virtual machines strengthen tenant isolation, while priority management protects real-time inference from background workloads.
2. What this changes for a Belgian or French company
An SME or mid-market organization can now frame requirements around concrete evidence: pre-production tests, environment separation, usage by team, call traces, and rollback procedures. For a large enterprise or public administration, tenant isolation can pool infrastructure without mixing data, quotas, or responsibilities.
IT teams must still distinguish product functionality from demonstrated compliance. A vendor score replaces neither a GDPR or AI Act risk assessment nor tests using the organization's real data, languages, and scenarios. Hosting location, administration, subprocessors, and reversibility still require contractual verification.
3. Underside analysis: industrializing RAG, agents, and Odoo
Underside's analysis is that observability must connect three levels: resources consumed, the agent's decision, and the business outcome. GPU and token dashboards control cost, but teams must also trace RAG documents, tool calls, permissions, human approvals, and the result in the target system.
For Odoo Enterprise, a sales, support, or accounting agent therefore needs a separate identity, least-privilege access, and per-operation limits. AutoRAG and agent templates can accelerate integration, but sovereignty depends on the entire chain—data, models, compute, logs, and operations—whether local, in a European cloud, or hybrid according to risk.
4. Operational recommendation
Before expanding a pilot, define a service owner, SLOs, budget, evaluation set, access tiers, and stop thresholds. Test updates on limited traffic, measure cost and quality by use case, and retain the evidence needed for audit without unnecessarily accumulating sensitive data.
Concrete priority: require a control sheet for every agent covering data, identity, tools, evaluations, cost, human supervision, rollback, and portability before connecting it to an ERP or business API.
Frame a governed AI platform